Scholarlink

<p>Anomalous Sound Detection (ASD) has research significance and application prospect industrial automation. Most existing models of ASD have limited ability to effectively utilize machine sound features, leading to reduced stability against sound anomalies and domain shift variations. To address the above issues, we propose a self-supervised ASD model based on spectrogram decomposition and parallel sub-network in this paper. Firstly, we decompose the spectrogram along the time and frequency dimensions to balance feature size and information integrity. This approach emphasizes the temporal and frequency variations in the feature map, facilitating a better understanding of the factors that affect machine sounds under domain shift conditions. Secondly, we design a pair of parallel training sub-networks. The parallel sub-networks employ self-attention mechanisms and shared gradients to effectively capture changes in features across both time and frequency dimensions. This approach improves model stability against anomalies and domain shifts. Finally, the anomaly scores of sub-network branches are fused as anomalous detection results. The performance of the proposed model is validated on DCASE2022 Task2 dataset. The Area under the Receiver Operating Characteristic Curve (AUC) and partial AUC (pAUC) of our model reached 72.89% and 64.83%. The results confirm the effectiveness of the proposed model, achieving better performance.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A self-supervised anomalous machine sound detection model based on spectrogram decomposition and parallel sub-network

  • 期刊论文

  • 发表日期：

    2025年3月1日

查看全文
  • Tao Zhang,
  • Lingguo Kong,
  • Xin Zhao,
  • Donglei Li,
  • Yanzhang Geng,
  • Biyun Ding,
  • Chao Wang

摘要

Anomalous Sound Detection (ASD) has research significance and application prospect industrial automation. Most existing models of ASD have limited ability to effectively utilize machine sound features, leading to reduced stability against sound anomalies and domain shift variations. To address the above issues, we propose a self-supervised ASD model based on spectrogram decomposition and parallel sub-network in this paper. Firstly, we decompose the spectrogram along the time and frequency dimensions to balance feature size and information integrity. This approach emphasizes the temporal and frequency variations in the feature map, facilitating a better understanding of the factors that affect machine sounds under domain shift conditions. Secondly, we design a pair of parallel training sub-networks. The parallel sub-networks employ self-attention mechanisms and shared gradients to effectively capture changes in features across both time and frequency dimensions. This approach improves model stability against anomalies and domain shifts. Finally, the anomaly scores of sub-network branches are fused as anomalous detection results. The performance of the proposed model is validated on DCASE2022 Task2 dataset. The Area under the Receiver Operating Characteristic Curve (AUC) and partial AUC (pAUC) of our model reached 72.89% and 64.83%. The results confirm the effectiveness of the proposed model, achieving better performance.